An adaptive framework for end to end quality of service management

In contrast to traditional text and data applications which are burst and elastic in nature, these emerging real-time multimedia applications are demanding on system resources such as ba

AN ADAPTIVE FRAMEWORK FOR END-TO-END

QUALITY OF SERVICE MANAGEMENT

LIFENG ZHOU

(B.S and M Eng., Nanjing University)

A THESIS SUBMITTED FOR THE DEGREE OF DOCTOR OF PHILOSOPHY DEPARTMENT OF COMPUTER SCIENCE NATIONAL UNIVERSITY OF SINGAPORE

2008

Acknowledgement

First and foremost, I wish to express my deepest gratitude to my supervisor, Dr Pung

Hung Keng and co-supervisor Dr Ngoh Lek Heng for their invaluable guidance and

support throughout my research efforts towards this thesis Their insights and

suggestions to the problems in this thesis have enlightened me in various detailed

aspects throughout the work

Dr Ooi Wei Tsang, Dr Samarjit Chakraborty and Associate Professor Roger

Zimmermann have served as my reviewers at different stages of this thesis I would

like to express my appreciation for their suggestions and comments and their time in

reviewing this thesis

I would like to thank all my colleagues in the Network Systems & Services (NSS)

Laboratory Among them, special thanks go to Dr He Jun, Dr Gu Tao, Dr Long Fei,

Dr Chen Lei, Ms Feng Yuan, Ms An Liming and Mr Suthon Sae-Whong for their

constant assistance and encouragements

Last but not least, I would like to thank my parents and my wife for their love,

unconditional support and patience during the course of my doctoral studies

Table of Contents

CHAPTER 1 INTRODUCTION 1

1.1 MOTIVATION 3

1.2 PROBLEM STATEMENT 6

1.3 THESIS CONTRIBUTIONS 7

1.4 THESIS OUTLINE 9

CHAPTER 2 LITERATURE REVIEW 10

2.1 QOS IN COMMUNICATION SYSTEMS 11

2.2 QOS PROVISIONING ARCHITECTURES 12

2.2.1 Network QoS Models 12

2.2.2 QoS-aware Operating Systems 14

2.2.3 QoS Middleware 15

2.2.4 Multimedia Applications and Media Framework 19

2.2.5 Cross-layer QoS Architectures 20

2.2.6 End-to-end QoS Schemes 22

2.3 DYNAMIC PROTOCOL COMPOSITION 23

2.4 SUMMARY 25

CHAPTER 3 THE QOS COORDINATION AND MANAGEMENT FRAMEWORK 26

3.1 REFERENCE MODEL FOR QOS MANAGEMENT 26

3.2 QCMF MANAGEMENT ARCHITECTURE 30

3.3 QCMF MANAGEMENT FUNCTIONS 32

3.4 SUMMARY 35

CHAPTER 4 END-TO-END QOS KNOWLEDGE MODELING 36

4.1 QOS KNOWLEDGE AND QOS ONTOLOGY 36

4.1.1 Related Work 36

4.1.2 General QoS Knowledge 38

4.1.3 QoS Ontology and RDFS Schema 40

4.1.4 QoS Ontology Predicates 42

4.2 APPLICATION QOS KNOWLEDGE MODELING 44

4.2.1 Motivation and Design Considerations 44

4.2.2 Two Layer Application QoS Ontology Model 48

4.2.3 QoS Domain Specification and Knowledge Acquisition 50

4.2.4 QoS Compilation and Mapping 55

4.3 MIDDLEWARE QOS KNOWLEDGE MODELING 57

4.3.1 Design Considerations 57

4.3.2 Ontology Modeling of Protocols 61

4.3.3 Semantic Protocol Stack Composition 64

4.4 NETWORK QOS KNOWLEDGE MODELING BRIEFING 67

4.5 QOS KNOWLEDGE PROCESSING 68

4.5.1 Knowledge Sharing 69

4.5.2 Knowledge Reasoning 71

4.6 SUMMARY 75

CHAPTER 5 END-TO-END QOS VIOLATION ANALYSIS 76

5.1 DESIGN CONSIDERATIONS 76

5.2 OVERVIEW OF OUR APPROACH 80

5.3 END-TO-END QOS VIOLATION ANALYSIS 83

5.3.1 End-to-end Monitoring of QoS Violations 83

5.3.2 Application QoS Violation Indicator 84

5.3.3 Correlate Application QoS Violations with Flow Statistics 86

5.4 VIOLATION CLASSIFICATION WITH NEURAL NETWORK 89

5.4.1 Neural Network Algorithms Briefing 90

5.4.2 Offline Algorithms 92

5.4.3 Online Algorithms 92

5.5 SUMMARY 94

CHAPTER 6 CROSS-COMPONENT QOS ADAPTATION 95

6.1 END-TO-END QOS MODEL 95

6.2 NETWORK QOS MODEL 98

6.3 END-HOST QOS MODEL 101

6.4 END-TO-END COORDINATION AND ADAPTATION 103

6.4.1 Information Gathering Algorithm 104

6.4.2 Cross-component Adaptation Evaluation Algorithm 106

6.4.3 End-to-end Signaling and Adaptation Algorithm 113

6.5 SIMULATION RESULTS 114

6.6 SUMMARY 119

CHAPTER 7 IMPLEMENTATION AND EVALUATIONS 121

7.1 IMPLEMENTATION SCENARIO 121

7.2 QOS KNOWLEDGE PROCESSING 123

7.2.1 SQS Initiation Delay 124

7.2.2 Knowledge Reasoning Performance 125

7.3 QOS VIOLATION ANALYSIS 127

7.3.1 Testing Cases 129

7.3.2 Data Analysis 134

7.4 END-TO-END QOS MANAGEMENT 140

7.4.1 QCMF Management Procedures 140

7.4.2 QCMF Management Performance 144

7.4.3 Control Channel Overhead 149

7.5 SUMMARY 151

CHAPTER 8 CONCLUSIONS AND FUTURE WORK 152

8.1 THESIS SUMMARY 152

8.2 FUTURE WORK 155

APPENDIX A ORTHONORMAL NETWORK FOR CLASSIFICATION 158

A SINGLE HIDDEN LAYER FEEDFORWARD NETWORK WITH RANDOM HIDDEN NODES 158

B APPROXIMATION WITH ORTHONORMAL BASIS 160

C GRAM-SCHMIDT ORTHONORMALIZATION 162

D SUMMARY OF ORTHONORMAL TRANSFORMATION 164

APPENDIX B AN EXAMPLE ONTOLOGY FOR PROTOCOLS 166

A PROTOCOL.RDF 166

B INSTANCE.RDF 167

BIBLOGRAPHY……… …170

List of Tables

TABLE 4-1: QOS PROFILES FOR MOBILE MULTIMEDIA APPLICATIONS 53

TABLE 4-2: PARTIAL RDFS REASONING RULE SET IN QCMF 72

TABLE 4-3: EXAMPLE FIRST-ORDER LOGIC RULES FOR COORDINATED QOS ADAPTATION 74

TABLE 5-1: TUNABLE PARAMETERS IN VIDEO TRANSMISSION, APPLICATIONS 84

TABLE 5-2: FLOW DESCRIPTORS FOR END-TO-END QOS 86

TABLE 6-1: SERVICE OPTIONS TABLE OF A NETWORK QOS COMPONENT 98

TABLE 6-2: SERVICE STATUS TABLE OF A NETWORK QOS COMPONENT AS IS MAINTAINED BY QMAN MIDDLEWARE INSIDE THE FLOW RECEIVER; FOR EACH NETWORK QOS COMPONENT, A CORRESPONDING TABLE IS KEPT BY QMAN AND UPDATED THROUGH EITHER PUSH OR PULL MODE 103

TABLE 6-3: SERVICE SUBSCRIPTION SETTINGS OF A FLOW IN SIMULATION AND THE ITS UTILITY FACTOR 114

TABLE 7-1: TESTBED CONFIGURATIONS 123

TABLE 7-2: QOS VIOLATION CLASSIFICATION IN VIEW OF CONTROLLABLE RESOURCES AND AVAILABLE END-TO-END ADAPTATION CHOICES 128

TABLE 7-3: QOS VIOLATION TEST WITH PLANETLAB NODES (SOURCE FROM NUS) 134

TABLE 7-4: SPECIFICATION OF QOS VIOLATION DATASETS: THE WIRED-LINE CATEGORY CONTAINS DATA OBTAINED FROM TESTBED, CAMPUS NETWORK AND PLANETLAB PLATFORM 134

TABLE 7-5: CLASSIFICATION ACCURACY OF QOS VIOLATIONS IN DIFFERENT ALGORITHMS 135

TABLE 7-6: CLASSIFICATION ACCURACY FOR QOS VIOLATIONS IN OUR ORTHONORMAL ALGORITHM 138

TABLE 7-7: USER-DEFINED ADAPTATION POLICIES FOR VIDEO STREAMING 144

TABLE 7-8: TIME TAKEN IN END-TO-END QOS MANAGEMENT 145

List of Figures

FIGURE 3-1: REFERENCE MODEL FOR END-TO-END QOS PROVISIONING AND COORDINATION 26

FIGURE 3-2: END-TO-END QOS TRANSMISSION SCENARIO 30

FIGURE 3-3: QCMF INCORPORATES BOTH HOST ARCHITECTURES AND NETWORK ARCHITECTURES 31

FIGURE 3-4: QCMF DESIGN CONCEPTS: CONTROL PLANE FOR SIGNALING, DATA PLANE FOR MEDIA TRANSMISSION AND KNOWLEDGE PLANE FOR META-DATA RECORDING 32

FIGURE 3-5: MANAGEMENT FUNCTIONS OF QCMF ARE FULFILLED BY ITS SEVERAL BUILD-TIME AND RUNTIME EXECUTION MODULES: SEMANTIC QOS SPECIFICATION (SQS) FOR KNOWLEDGE MODELING, MIDDLEWARE QOS MANAGER (QMAN) FOR RUNTIME MANAGEMENT AND DYNAMIC PROTOCOL FRAMEWORK (DPF) FOR MIDDLEWARE LEVEL ADAPTATION 35

FIGURE 4-1: PARTIAL QOS ONTOLOGY FOR ACCESS NETWORK WRITTEN IN RDFS 43

FIGURE 4-2: SEMANTIC MODELING AND SYNTACTICAL QOS SPECIFICATION IN QCMF 47

FIGURE 4-3: THE HIERARCHICAL APPLICATION QOS ONTOLOGY MODEL 48

FIGURE 4-4: PARTIAL QOS DOMAIN SPECIFICATION FOR VIDEO STREAMING APPLICATIONS 51

FIGURE 4-5: AN EXAMPLE OF KNOWLEDGE BUILT IN THE VIDEO-AUDITORY QOS DOMAIN 52

FIGURE 4-6: DYNAMIC COMPILATION OF AQOSPEC 55

FIGURE 4-7: ARCHITECTURE OF DPF WITH ONTOLOGY MODELING 61

FIGURE 4-8: PROTOCOL KNOWLEDGE MODELING ENTRY POINT: SERVICE AND CATEGORY CLASSES 62

FIGURE 4-9: TCP IS OF (RDF:) TYPE TRANSPORT AND BELONGS TO TRANSPORT CATEGORY 63

FIGURE 4-10: SEMANTIC PROTOCOL SELECTION AND PROTOCOL STACK BUILDING 65

FIGURE 4-11: RDFS DEFINITION FOR COMPATIBILITY AND DEPENDENCY 66

FIGURE 4-12: END-TO-END QOS KNOWLEDGE SHARING AND ADAPTATION SIGNALING 69

FIGURE 4-13: ONTOLOGY DEFINITIONS FOR SOME OS TYPES AND INSTANCES INFORMATION 73

FIGURE 5-1: OBSERVED JITTER VARIATION IN A VIDEO TRANSMISSION 79

FIGURE 5-2: SINGLE HIDDEN LAYER FEEDFORWARD NEURAL NETWORKS 90

FIGURE 6-1: ABSTRACTED END-TO-END QOS PROVISIONING MODEL 95

FIGURE 6-2: END-HOST QOS MANAGEMENT MODEL (QMAN) 101

FIGURE 6-3: SKELETON OF THE CROSS-COMPONENT ADAPTATION EVALUATION ALGORITHM 107

FIGURE 6-4: DELAY CHANGE AT NETWORK QOS COMPONENT 2 WHERE A VIOLATION HAPPENS AND COMPONENT 4 WHICH PARTICIPATES IN THE END-TO-END COLLABORATION TO SOLVE THE VIOLATION 118

FIGURE 6-5: EXPERIENCED END-TO-END DELAY BEFORE/AFTER A DELAY VIOLATION 119

FIGURE 6-6: DELAY OVERHEAD OF ADAPTATION ALGORITHMS AEA AND ASU FOR MESSAGE EXCHANGE AND SIGNALING AMONG NETWORK QOS COMPONENTS AND END-HOSTS 119

FIGURE 7-1: TESTBED ENVIRONMENTS 122

FIGURE 7-2: OVERHEAD OF THE TWO-LAYER ONTOLOGY DESIGN 124

FIGURE 7-3: THE ONTOLOGY REASONING PERFORMANCE 126

FIGURE 7-4: KNOWLEDGE REASONING PERFORMANCE COMPARISON 127

FIGURE 7-5: CPU OCCUPIER PROGRAM FOR CPU VIOLATION AT END-HOSTS 128

FIGURE 7-6: OBSERVATION OF END-TO-END QOS W/ AND W/O CPU CONTENTION 130

FIGURE 7-7: TRAFFIC GENERATOR CAN PRODUCE TRAFFIC OF EITHER CONSTANT RATE OR NORMAL DISTRIBUTION 131

FIGURE 7-8: OBSERVATION OF END-TO-END QOS W/ AND W/O NETWORK CONGESTION 131

FIGURE 7-9: OBSERVATION OF END-TO-END QOS VARIATION IN WIRELESS COMMUNICATION 133

FIGURE 7-10: TESTING CLASSIFICATION ACCURACY COMPARISON BETWEEN LMBP AND ELM 137

FIGURE 7-11: TRAINING TIME COMPARISON BETWEEN LMBP AND ELM 137

FIGURE 7-12: PERFORMANCE OF THE PROPOSED ORTHONORMAL ALGORITHM IN QOS VIOLATION CLASSIFICATION: (A) TRAINING AND TESTING ACCURACY CURVES, (B) TRAINING TIME CURVE 138

FIGURE 7-13: AN EXAMPLE QLIST FOR VIDEO STREAMING 140

FIGURE 7-14: GRAPHIC USER INTERFACE (GUI) FOR STREAMING 142

FIGURE 7-15: NETQ PROGRAM FOR DATA PACKET CAPTURING AT THE MEDIA RECEIVER 143

FIGURE 7-16: SAMPLE SPARQL QUERY FOR ONTOLOGY INTEGRITY CHECK BETWEEN TWO INSTANCE CLASSES 145

FIGURE 7-17: STREAM DELIVERY AND ADAPTATION AT THE MEDIA SENDER 147

FIGURE 7-18: STREAM RECEIPT AND ADAPTATION AT THE MEDIA RECEIVER 147

FIGURE 7-19: QUALITY FLUCTUATION OF THE RECEIVING FRAME RATE BEFORE, DURING AND AFTER CONGESTION VIOLATION 148

FIGURE 7-20: END-TO-END FLOW STATISTICS OF AN AUDIO STREAMING UNDER VIOLATION 149

FIGURE 7-21: RMI INVOCATION DELAY FOR THE CONTROL PLANE (LOGARITHM SCALE) 150

Summary

High-speed networks and powerful end-hosts enable new types of Quality of Service

(QoS) sensitive applications such as Video-On-Demand to be offered In contrast to

traditional text and data applications which are burst and elastic in nature, these

emerging real-time multimedia applications are demanding on system resources such

as bandwidth and CPU, and are also sensitive to continuous QoS performance To

provide end-to-end QoS to users, researchers have spent great efforts in finding

suitable QoS provisioning mechanisms in areas such as QoS middleware, adaptive

applications and QoS-aware networks We find that the approaches of most existing

researches have been piecemeal, wherein each focusing on a different aspect of the

QoS provisioning mechanisms We argue that the real design issue of end-to-end QoS

is more complex than when each of these QoS mechanisms is considered on its own It

is therefore not sufficient to rely merely on, say middleware, applications or networks

to fulfill end-to-end QoS Instead, an integrated approach to the overall end-to-end

QoS provisioning, harmonizing QoS mechanisms in the applications, middleware and

networks are essential

In this thesis, we propose an adaptive end-to-end QoS coordination and management

framework (QCMF) for the QoS management of multimedia applications Unlike other

end-to-end QoS architectures which mainly focus on the interface design between

adjacent layers, resource reservation or work-flow management, QCMF aims at

designing an effective end-to-end QoS platform for accommodating and coordinating

QoS efforts from heterogeneous end-to-end QoS components (e.g., end-host QoS

management and network QoS provision) Our solution encompasses existing or new

QoS mechanisms at three levels: the network level, the middleware level and the

application level, each of which is abstracted as a meta-model in the end-to-end QoS

scenario where their behaviors and interactions are studied The proposed framework

is adaptive in the sense that it recognizes and coordinates the adaptive behaviors of

multimedia applications and networks in view of the changing runtime environment

context Besides, QCMF provides the ability of dynamic composition of end-hosts’

communication stacks, which provides another possible dimension of QoS adaptation

at the middleware level

With the aforementioned methodology in mind, we have proposed a set of techniques

to fulfill our overall design objectives of a coordinated end-to-end QoS management

Firstly, we propose a unified knowledge plane for end-to-end QoS modeling, in which

QoS information of each end-to-end QoS component is described semantically The

semantic approach of modeling QoS knowledge facilitates the deployment of

multimedia applications in heterogeneous environments where services of desirable (or

compatible) features can be selected according to runtime service availability

Moreover, information sharing among QoS components becomes easier as different

end-to-end QoS components would have a common understanding of QoS knowledge

while interacting with each other Secondly, we propose a novel approach to the

analysis of QoS violations By monitoring end-to-end flow statistics and application

performance, a QoS violation can be quickly identified with high accuracy Such an

approach outperforms traditional rule-based violation detection methods which have

seldom undergone a rigorous testing procedure and require clear margins of QoS

parameters in asserting a QoS violation Lastly, we propose an end-to-end QoS

coordination scheme and algorithms for runtime collaborative end-to-end QoS

management By exchanging QoS information and coordinating adaptation behaviors

among QoS components, a QoS violation can be solved by either a local adjustment at

the QoS component where the violation takes place or being processed by another QoS

component participating in the end-to-end collaboration Such a decision is made at

end-hosts in a pure end-to-end fashion without violating the end-to-end design

principle of the Internet Our prototype implementation validates our design

philosophy and demonstrates that QCMF is functional Performance evaluation results

of the prototype show that QCMF works effectively in many aspects of end-to-end

QoS management such as control signaling, knowledge processing, violation detection

and coordinated adaptation

Publications

[1] Lifeng Zhou, Lei Chen, Hung Keng Pung and Lek Heng Ngoh, “Identify QoS

Violations through Statistical End-to-end Analysis”, to submit to a journal, 2009

[2] Lei Chen, Lifeng Zhou and Hung Keng Pung, “Universal Approximation and

QoS Violation Applications of Extreme Learning Machine”, accepted by Neural

Processing Letters, Springer, 2008

[3] Lei Chen, Lifeng Zhou and Hung Keng Pung, “Universal Approximation

Analysis and Applications of Orthonormal Neural Networks”, submitted to

International Journal of Pattern Recognition and Artificial Intelligence, 2008

[4] Lifeng Zhou, Lei Chen, Hung Keng Pung and Lek Heng Ngoh, "End-to-end

Diagnosis of QoS Violations using Neural Networks", in Proc the 33rd IEEE

International Conference on Local Computer Networks (LCN), 2008

[5] Lifeng Zhou, Hung Keng Pung and Lek Heng Ngoh, “An end-to-end Framework

for Coordinated QoS Adaptation”, in Proc the 33rd IEEE International

Conference on Local Computer Networks (LCN), 2008

[6] Lei Chen, Lifeng Zhou and Hung Keng Pung, “Approximation Capability of

Feedforward Neural Networks with Least Square Solutions”, accept with major

revision by NeuroComputing, 2008

[7] Lifeng Zhou, Hung Keng Pung, Lek Heng Ngoh and Tao Gu, “Ontology

Modeling of a Dynamic Protocol Stack”, in Proc the 31st IEEE International

Conference on Local Computer Networks (LCN), 2006

[8] Lifeng Zhou, Hung Keng Pung and Lek Heng Ngoh, “Towards Semantic

Modeling for QoS Specification”, in Proc the 31st IEEE International

Conference on Local Computer Networks (LCN), 2006

[9] Lifeng Zhou, Hung Keng Pung and Lek Heng Ngoh, “Knowledge Modeling for

End-to-End QoS Management”, in Proc the 1st International Conference on

Communications and Networking in China (ChinaCom), 2006

[10] Liming An, Hung Keng Pung and Lifeng Zhou, “Design and Implementation of

a Dynamic Protocol Framework”, Computer Communications, Volume 29, Issue

9, pp 1309-1315, May 2006

[11] Suthon Sae-whong, Hung Keng Pung and Lifeng Zhou, “QMan: an Adaptive

End-to-End QoS Architecture”, in Proc the 12th IEEE International Conference

on Networks (ICON), 2004

Chapter 1 Introduction

I N T R O D U C T I O N 1

Most conventional and legacy network applications have been designed to operate in

low to moderate speed (a few 10s of Mbps) network environments These networks

have been useful and adequate for supporting text and data applications, including

distributed applications requiring short requests-responses, in which relatively small

amount of bandwidth is needed for each transmission Moreover, these applications are

rather elastic in nature, i.e., they can tolerate great variations in performance such as

packet delay and throughput rates In recent years, great advance has been made in

communication technologies where networks that can support data traffic in gigabits

per second on every port (e.g., Gigabit Ethernet) are now available off-the-shelf As a

result, high-speed networks and powerful end-hosts enable new types of applications

such as Video-on-Demand, multimedia-based collaborative computing and

teleconferencing In contrast to traditional elastic data applications, these emerging

multimedia applications have different traffic characteristics and are demanding on

system resources such as network bandwidth and CPU time slice The challenges in

designing such applications generally lie in catering for time dependent (or continuous)

media: audio and video Besides storage speed, memory size and processing power,

timely delivery of media data over networks is also an essential factor It requires not

only considerable computing resources, but also ensures that these resources will be

available over a certain period of time Failure to sustain such provisioning will

generally compromise the presentation quality of continuous media Thus, the need for

transmission quality assurances for these Quality of Service (QoS) sensitive

applications arises naturally

The question of which suitable mechanisms for the provisioning of QoS has been

asked and possible answers suggested and issues debated One interesting but trivial

suggestion of solution is through over-provisioning of bandwidth The motivation of

this thought is that bandwidth, due to increasing availability of fibers and

wavelength-division multiplexing (WDM) technique, is potentially abundant and cheap We

believe over-provisioning can greatly ease QoS problems but is not a panacea This is

because of at least three main reasons: (1) not all QoS problems are constrained by

bandwidth, jitter is a classical example; (2) no matter how much bandwidth the

network can provide, new innovative applications is likely to be created in the near

future to consume them [1]; (3) unless there is a common physical transmission

technology (fiber is a potential candidate) for all different network solutions, the vision

of the abundant bandwidth cannot be materialized for a very simple reason: all

networks are to be interconnected in one way or the other and hence those networks of

lower bandwidth will become the QoS bottleneck Indeed LANs, dial-ups, wireless

LANs, WANs and broadband co-exist and interconnect to form the global Internet

The use of high-speed core networks has not eliminated QoS problems, as we have

known and experienced today For example, wireless communication technologies,

including wireless LAN (IEEE802.11b/g), Bluetooth and 3G mobile networks, are

being developed and deployed as common services nowadays, which enables wireless

multimedia streaming to be delivered in light-weighted devices such as handset As

end-applications are very likely to run over either fixed wired networks or wireless

networks, the overall network environment becomes more dynamic and heterogeneous

Hence the QoS problems are more difficult to be resolved by relying on a simple

mechanism, such as over-provisioning of bandwidth Suitable QoS mechanisms are

needed in networks and end-hosts to best assure the timely delivery of multimedia data

For over ten years, researchers have proposed various QoS solutions in either

end-hosts or networks In QoS provisioning through networks, researches have been

focused on providing suitable QoS models and service disciplines, as well as

appropriate admission control and resource reservation protocols For example, the

Internet Engineering Task Force (IETF) has defined several standard QoS

architectures such as Integrated Services (IntServ) [2], Differentiated Services

(DiffServ) [3], Constraint-based Routing [4] and Multiprotocol Label Switching

(MPLS) [5] IntServ and DiffServ are well-known network QoS models, which have

been studied and compared (through simulation, prototyping and performance

measurements) by many researchers IntServ, relying on the Resource Reservation

Protocol (RSVP) [6], duels with resource allocations and reservations for each data

flow and hence would have the potential to provide guaranteed QoS service Many

network vendors, such as Cisco and Sun Microsystems have IntServ/RSVP

implementations on their routers [7] On the other hand, DiffServ is based on a simple

model where traffic entering a network is aggregated into classes and treated

differently within a DiffServ-enabled network There are router prototypes and

products actually implementing DiffServ service MPLS is a forwarding scheme that

has the ability to aggregate traffic flows and hence can provide a basis for both IntServ

and DiffServ QoS support over core networks Constraint-based routing intends to

address QoS from the routing point of view by establishing an appropriate route

meeting some QoS constraints such as bandwidth or/and delay requirements

In end-hosts, various QoS architectures have been proposed and discussed According

to their resource management styles, these solutions can be categorized into

reservation-based approaches and adaptation-based approaches [8]

Reservation-based approaches employ resource reservation and admission control mechanisms

(such as CPU preemption and scheduling) to guarantee the availability of resources

before multimedia data is delivered [9] The sustaining of transmission quality depends

on the QoS technologies of the underlying platform (e.g., QoS capability of the

operating system and network), in which the data packets are handled Nevertheless,

because of the following reasons, multimedia transmission cannot rely solely on such

resource allocation and reservation mechanisms

• QoS degradation in best effort networks is often unavoidable [10], as QoS assurance provided by the underlying services may vary from time to time

• The Internet traffic produced by end-users exhibits a dynamic behavior There has been no effective QoS reservation mechanism for dueling with the diverse

QoS requirements of applications and the dynamic behavior of the network

traffic

• QoS guaranteed technologies have yet to be established as common services, hence most today’s networks are still operating in best-effort or best assured

mode

In view of the above restrictions, QoS adaptation, which allows a multimedia

application to react suitably to occurring QoS violations, is essential to ensure that the

application can sustain certain level of QoS in various runtime environments An

adaptation-based QoS approach can operate in best-effort or QoS-enabled network

environments and manages QoS in a pure end-to-end fashion where QoS monitoring,

analysis and adaptation are enforced throughout the lifecycle of the transmission to

smooth the quality fluctuation and best maintain the agreed QoS level An

adaptation-based approach requires minimum modification to existing network architecture, thus

makes itself more suitable to be deployed over current non-real-time OS and

best-effort network environments (or future QoS-enabled network environments)

In adaptation-based QoS researches, progresses have been made in several directions

such as QoS-aware applications, QoS middleware or QoS-enabled operating systems

Most work done in the application layer is related to the transmission of continuous

media streams (e.g variable bit rate codec, media compression, frame-dropping and

layered encoding scheme), and hence is rather media specific and restrictive in certain

application domains [11][12][13][14][15] On the other hand, researchers have also

proposed research prototypes of QoS enabled/sensitive operating systems, applying

results from real-time scheduling theory to support system level QoS management

[16][17][18] However, such an approach would often result in a proprietary OS,

which is therefore not popular In recognition of these limitations, more active research

efforts have been devoted to provide QoS supports as middleware services

[19][20][21][22][23][24] The QoS middleware approach is popular for at least two

main reasons despite of its performance overhead: (1) the QoS solutions are likely to

be independent of the network and OS platforms, and (2) the QoS controls can be

specifically designed and possibly be transparent to applications

This thesis proposes an adaptive end-to-end QoS Coordination and Management

Framework (which we call QCMF) for QoS management of end-to-end multimedia

transmission Different from most existing work that focuses on a particular QoS

provisioning domain (e.g., networks or applications), QCMF provides an integrated

solution for end-to-end QoS management which designs a set of techniques to embrace

existing or new QoS efforts from different areas of end-to-end provisioning Details of

our approach will be discussed later

The need to provide QoS support for networked multimedia applications has long been

recognized and discussed As QoS issue has not been part of the design considerations

of virtually all network architectures, including that of the Internet, the design and

development of suitable QoS provisioning mechanisms has to be carefully considered

so as to ensure the stability of current Internet architecture and its compatibility with

other add-on network services such as Network Address Translation (NAT) [27] In

fact, the complexity of QoS provisioning has already resulted in various QoS solutions

each focusing on a different aspect of the QoS provisioning mechanisms, depending on

the perspectives and design centric of the designers As discussed, these solutions can

be broadly classified into three main design viewpoints: QoS-aware applications,

dedicated QoS middleware and network QoS models

However, the real design issues of QoS provision are far more complex than when

each of these design viewpoints is considered on its own This is simply because

meeting performance requirements of QoS-sensitive applications is fundamentally an

end-to-end issue It requires all QoS-enabled facilities along the end-to-end path

working cohesively to achieve the desired end-to-end performance As most existing

QoS solutions focus on their respective areas while paying little attention to the

interaction with other QoS services on the end-to-end path, QoS can only be sustained

in their local domains, while no satisfactory end-to-end performance can be provided

to users In this sense, we believe that a more holistic approach to the overall

end-to-end QoS provisioning, integrating QoS mechanisms in the applications, middleware

and the networks is essential

An adaptive QoS coordination and management framework (QCMF) has been

developed based on such a design consideration The framework embraces QoS

services along the provisioning path and provides mechanisms for QoS coordination

and adaptation among them in both build-time instantiation and runtime QoS

management Different from existing integrated end-to-end architectures which have

typically developed a whole set of new end-to-end QoS mechanisms by themselves

[9][28], QCMF aims at accommodating existing QoS techniques from different

domains and providing a platform for their interaction For instance, QCMF does not

invent any new signaling protocol for QoS negotiation among end-to-end QoS

components (opposite to [29]), but makes use of any existing protocols capable of

negotiation Unlike [30] which designs its own network QoS implementation as part of

its end-to-end QoS efforts, QCMF assumes a generic network service differentiation

model for end-to-end collaboration Such a model can be easily mapped to existing

standard network QoS models such as DiffServ which is built on the same basic QoS

discipline of service differentiation In this way, QCMF requires minimum

modification of current network architecture and hence has a better chance to be

accepted and implemented as common utility services over the Internet

1.3 TH E S I S CO N T R I B U T I O N S

This thesis proposes an adaptive QCMF framework for QoS management in

end-to-end multimedia transmission The solution embraces existing and new QoS

mechanisms at three entity levels: networks, middleware and applications QCMF

provides necessary management functions that include, for example, QoS negotiation,

monitoring and adaptation With runtime adaptation, QCMF enables multimedia

applications to best maintain certain degree of QoS under constrained system and

network resource availability In summary, this thesis makes the following key

contributions:

1 We propose a new design philosophy with respect to how current communication

architectures of end-hosts and networks could be modified to accommodate

end-to-end QoS services Rather than designing a new set of QoS mechanisms for each

communication layer so that they can be seamlessly integrated together for

end-to-end QoS provisioning, we propose to unify existing isolated QoS solutions at

different layers so as to fulfill end-to-end QoS requirement We believe our

solution is easier to be implemented and deployed in current network environment

2 (As the continuation of point 1) we propose a set of techniques to enable the

collaboration among end-to-end QoS systems We treat each of the QoS

sub-systems as a meta-component and design an end-to-end framework and methods

for accommodating and supporting interactions and dynamic adaptations among

them In this context, we are not participating in the performance enhancement of

QoS mechanisms of any individual layer Instead, our contribution is to provide a

platform for harmonizing and coordinating existing QoS mechanisms in

applications, middleware and networks in the context of overall end-to-end QoS

provisioning

3 We propose a uniform semantic approach and meta-models to abstract QoS

characteristics of applications, middleware and networks Each of these

meta-models will provide consistent interfaces so as to facilitate interactions among

adjacent QoS models Based on such a semantic specification method, we establish

a knowledge plane for QoS information exchange among different

QoS-subsystems for the benefit of end-to-end QoS negotiation and management The

advantages of such an approach lie in a powerful and expressive method for

specification as well as an easy way for information processing, matching and

sharing

4 We propose a novel end-to-end approach to QoS management with respect to the

diagnosis of QoS violations By monitoring end-to-end flow statistics and

application performance, a QoS violation can be quickly identified with high

accuracy as we have tested Such an approach outperforms a traditional rule-based

violation detection method which has seldom undergone a rigorous testing

procedure and requires clear threshold values of QoS parameters in asserting a QoS

violation

5 We demonstrate the design concepts of points 1-4 and the functionality of the

proposed QCMF framework through prototype implementation We have

developed a set of software reflection techniques for the implementation of

meta-QoS models for applications, middleware and networks In addition,

decision-making algorithms, heuristics and policies have been defined for a collaborative

end-to-end QoS management Through physical measurements of our

implementation, we have shown that QCMF can achieve the aforementioned

features and functionalities with acceptable overhead

1.4 TH E S I S OU T L I N E

The rest of the thesis is organized as follows

Chapter 2 surveys relevant literatures in areas of end-host QoS research and network

QoS research We discuss the features and limitations of existing approaches The

differences between QCMF and previous work are also compared

Chapter 3 gives an overall picture of QCMF by explaining its design philosophy (i.e.,

design reference model), system architecture and management functionalities

Chapter 4 elaborates the knowledge modeling in QCMF whereby characteristics of

each QoS sub-system with respect to end-to-end collaboration are semantically

abstracted and processed

Chapter 5 explains our approach for runtime QoS monitoring and violation analysis

We also give an overview of the violation identification algorithms we have engaged,

whose performances are compared and discussed in Chapter 7

Chapter 6 presents the cross-component adaptation scheme in QCMF Detailed

description about our design assumptions, meta-models for end-to-end QoS

components and coordination algorithms and heuristics are explored Simulation

models and results are then introduced which has validated the correctness of our

approach

Chapter 7 describes our prototype implementation and performance measurements of

QCMF

Chapter 8 concludes the thesis and discusses future work

Chapter 2 Literature Review

C HAPTER

L I T E R A T U R E R E V I E W 2

The open problem of QoS provisioning has been addressed by various research efforts

in the past years In this chapter, we will review some of the advance in both network

and end-host QoS researches More comprehensive end-to-end QoS solutions such as

cross-layer architectures and integrated end-to-end QoS systems will also be

introduced and compared By examining these related researches, we will show the

advantages of our work over previous studies

2.1 QOS I N CO M M U N I C A T I O N SY S T E M S

The term QoS is first introduced to describe characteristics of low-level data

transmission in communication systems With the appearance of distributed

multimedia applications, the meaning of QoS has been re-defined as “the collective

effect of service performance which determines the degree of satisfaction of a user of

the service” [31] In general, QoS represents a set of quantitative and qualitative

characteristics of a distributed multimedia system that are necessary to achieve the

required functionality and performance of an application Here functionality and

performance refers to both the proper delivery of media data to a multimedia

application user and the overall user satisfaction [32]

In practice, QoS is often expressed using measurable QoS parameters A QoS

parameter describes a specific attribute of a communication system or a performance

requirement of a multimedia application Each QoS parameter can be viewed as a

typed variable with bounded values An application’s QoS requirements are conveyed

in terms of high-level QoS parameters that specify what the application desires These

QoS parameters are assessed by the underlying communication system to determine

whether application requirements can be met or not

The underlying system needs resources to promote its service to multimedia

applications Essentially, there are two kinds of resources relevant to the performance

of a multimedia application: end-host resources and network resources The former

consists of processing power, memory, data buffer in an end-host and its peripheral

multimedia devices; the latter includes network bandwidth and packet queuing priority

To manage these resources for applications, two camps of QoS researches have been

established focusing on their respective areas, namely end-host QoS research and

network QoS research

2.2 QOS PR O V I S I O N I N G A R C H I T E C T U R E S

The open problem of providing end-to-end QoS support has been addressed by various

research efforts in the past years [9][28][32][33] This section reviews existing QoS

researches applicable to areas such as network QoS, end-host QoS and end-to-end QoS

2.2.1 Network QoS Models

To support QoS in the Internet, IETF has defined several standard service models and

mechanisms to meet the demand for QoS The IntServ/RSVP [2][6] architecture

intends to provide end-to-end bandwidth reservation by maintaining per-flow state

information along the path from the flow sender to the receiver However, the

complexity of per flow operations usually increases as a function of the number of

flows In addition, it is difficult to maintain the consistency of per flow state in a

distributed network environment Thus the IntServ model is not scalable to large

networks [1] Such a scalability problem has resulted in the DiffServ approach [3]

where QoS is achieved by a coarse level of service differentiation among a small

number of traffic classes The main advantage of DiffServ over IntServ is that core

network will only operate on aggregated flows instead of per flow in IntServ In edge

routers, packets are processed and aggregated on the basis of service classes However,

the DiffServ solution will become complex when QoS is to be offered over multiple

DiffServ domains Notably, there is a widely used QoS reference model merging these

technologies This includes the models combining IntServ in access network and

DiffServ in the backbone network [34] MPLS [5], on the other hand, is a layer two

forwarding scheme that has the ability to aggregate traffic flows and hence can provide

a basis for both IntServ and DiffServ QoS support over the core network

Network QoS research in recent years mostly focuses on (1) the functional

improvement of these standard QoS models through techniques such as traffic

engineering [35][36], or (2) discusses the impact of these models on existing

communication facilities such as the performance variation of TCP protocol [37]

Nevertheless, we should note that network QoS models or solutions discussed above

can only deliver end-point to end-point QoS, i.e., from the network egress point of a

flow sender to the ingress point of a flow receiver However, the main body of QoS

communication lies within both end-hosts and their applications In another word,

what we want to satisfy is the QoS requirements from multimedia applications, which

is more precisely, application-to-application QoS The network QoS models by

themselves, can not provide application-to-application QoS A simple example is that,

the fluent delivery of video frames to an end-user relies on network resources such as

bandwidth and end-host resources such as CPU time slice While a network QoS may

assure the provision of bandwidth, the successful end-to-end QoS provision still

depends on the sufficient CPU time slice allocated at both flow sender (for media

encoding) and receiver (for decoding) The gap between application-to-application

QoS and network QoS is left for end-host QoS to bridge Moreover, QoS is not always

fully guaranteed in these proposed network QoS models For instance, DiffServ

provides a sense of resource allocation and service differentiation, but it never

guarantees the provision of QoS in the network: intra service class bandwidth

contention in a DiffServ domain is often managed by traffic engineering technologies

such as statistical admission control [38] and Random Early Detection (RED) [39] It

is obvious that such traffic engineering technologies cannot strictly guarantee even

network-wide QoS Thus an end-to-end flow may expect temporary quality fluctuation

during transmission where end-host QoS mechanisms may take their places

2.2.2 QoS-aware Operating Systems

A number of pioneering efforts have produced useful QoS provisioning mechanisms in

end-hosts, among which QoS-aware operating system research has once been a focus

To support the execution of real-time multimedia applications, the operating system of

a computer has been argued to have the ability to manage and resolve resource

contentions of these applications so as to ensure timely processing and delivery of

multimedia data

Several research prototype operating systems have emerged, applying results from

real-time scheduling theory For example, the DASH kernel [16] uses an admission

control algorithm based on a timeline and then uses earliest deadline scheduling to

actually sort all tasks In order to guarantee the performance of an application,

computational requirements of the application need to be measured beforehand and be

analyzed together with its timing constraints such as delay bounds In this way, an

application can be executed within expectation where its timing constraints can be

satisfied Similar observation can be found in RT-Mach [17] and Pegasus [18] where

applications need to specify timing constraints explicitly Based on that information,

the OS kernel can calculate its CPU usage and provide fine-grained timestamp and

synchronization

In recent years, great strides have been made to support QoS provisioning in

commercial OS and network products Most Windows operating systems are now able

to signal RSVP and do kernel level packet scheduling [40] There are also several

add-ons available to win32 platforms which can provide advanced QoS supports such as

CPU resource reservation [41] On the other hand, large network vendors, such as

Cisco and Sun Microsystems have embedded DiffServ on their high end routers [42]

However, as the Internet today is still best effort, there is no means to reserve network

resources such as bandwidth, which is vital to end-to-end QoS provisioning Thus

these low level (OS and network) QoS supports are still tentative and premature in

nature

2.2.3 QoS Middleware

Traditionally, middleware is a layer of software that runs above heterogeneous

operating systems and communications systems, providing a uniform interface to

distributed applications In end-host QoS researches, various projects have been

proposed to provide QoS supports as middleware services Typically, a QoS

middleware provides services ranging from QoS specification, negotiation to runtime

supervision The following paragraphs will provide a detailed discussion on some of

the latest QoS middleware and compare their key features with those of our QCMF

DaCaPo++ [43] is a middleware QoS project that supports a range of multimedia

applications It automatically configures itself at start-up time to provide suitable

communication protocols and multimedia oriented services that are adaptable to

application needs MCF [20] from the same research group offers flexible multipoint

communication services through protocol configurations at start-up time To make

QoS parameters more application friendly, “types” of media can be specified in both

MCF and DaCaPo++ where different treatment will be provided to each media type

On the other hand, DJINN [24] and Chameleon [44][45] are designed to support

runtime protocol stack re-composition in addition to build-time composition, which

offers more flexibility of QoS adaptation than DaCaPo++ DJINN allows application

developers to create and connect model components in the form of connection

diagrams At runtime, such a component graph can be modified if intra-components

reconfiguration can not solve a QoS violation In a heavy loaded network environment,

for example, the congestion control mechanism of TCP may introduce unnecessary

overhead to a multimedia stream which can tolerate certain degree of packet loss

Through runtime re-composition, a TCP protocol component can be replaced with

light-weighted protocol such as UDP in DJINN Leveraging on the dynamic protocol

framework (DPF) [46] component, our QCMF provides similar build-time stack

composition and runtime re-composition compared with DJINN In the context of

QCMF, DPF offers a possible dimension of QoS adaptation at the middleware level

However, the QoS adaptation issue (e.g., end-to-end information sharing and

decision-making) in QCMF is more carefully designed compared with aforementioned

researches in that it also reviews adaptation choices in other domains such as

multimedia applications (e.g., variable video frame rate) and networks (e.g., service

class upgrade in DiffServ) In this sense, QCMF offers a more comprehensive

end-to-end solution where middleware level adaptation is only one of the runtime

considerations

The 2KQ project [47] from UIUC proposes a resource-aware service configuration model for heterogeneous distributed environments 2KQ employs multi-tie QoS translation Firstly, specification of the application is translated into a set of component

configurations Secondly, the set of component configurations are translated into

QoS-aware component specification (QoSCSpec) Lastly, QoSCSpec is translated into the

corresponding system QoS parameters and their resource requirements (e.g., CPU or

network bandwidth) The QoS specification and mapping process of QCMF is similar

to that of 2KQ However, QCMF proposes a systematic semantic model to describe the roles and relationships among various QoS entities including middleware components,

network QoS services and application requirements As a result, standard high level

QoS entities can be more easily matched and mapped into system level resource

specification

Agilos [22] is a middleware control architecture to assist application-aware adaptations

The main contribution of this project is to introduce a fuzzy control model for the

decision-making of QoS adaptations The correctness and efficiency of their model

have been proven by mathematical analysis and prototyping Agilos utilizes fuzzy

rules in the form of “IF-THEN-ELSE” clause to define adaptation behavior However,

as system complexity increases, reliable fuzzy rules and membership functions used to

describe system behavior are difficult to determine Comparatively, QCMF engages a

machine learning approach to QoS violation analysis By examining the end-to-end

flow statistics and application behavior, QCMF can identify a QoS violation without

the need to specify threshold values for communication parameters Moreover, Agilos

is specifically designed for those applications that receive control commands from the

middleware Hence, Agilos does not allow applications to specify fuzzy rules as

adaptation decision is solely made by analytical translation through middleware probe

service [8] A similar approach is taken in [48] which defines strategic and tactical

QoS managers Strategic QoS managers take a global view of QoS provided by a set of

application components within the manager’s policy domain while tactical QoS

managers provide local control over application components In contrast to these

studies, there is virtually no restriction on the kind of multimedia applications that

QCMF can serve For those applications that have their own QoS logics which are out

of the control of a QoS middleware, QCMF provide information support by

establishing a knowledge plane for information record and exchange (Chapter 4) For

other applications that do not have built-in intelligence for QoS management, QCMF

will guide the behavior of these applications through end-to-end coordination In both

scenarios, QCMF allows application-specific policies to be defined, which is used to

direct the management behavior and adaptation decision-making of the end-to-end

QoS system (Chapter 6)

Through reviewing these recent researches, we have identified the following trends in

the design of emerging QoS middleware Firstly, QoS middleware are becoming more

and more flexible Many QoS middleware today are designed in component-based

architectures, meaning that various functionalities are encapsulated into components

and can be swapped in and out on the fly [24][49][50] In this way, higher flexibility

can be achieved where customized services can be provided to a multimedia

application Secondly, several QoS middleware has incorporated additional features

such as multipoint and security support [23], which makes them more versatile in

supporting a wide range of application needs Lastly, more and more network

applications incorporate multimedia contents and require corresponding QoS supports

As a result, purposeful middleware has been proposed to serve a specific application in

a particular environment [51][52] For example, [53] has designed a distributed

middleware for networked audio and visual home appliances, which is executed on

commodity software Built on Linux platform, such a middleware can control a wide

range of home appliances

2.2.4 Multimedia Applications and Media Framework

As stated earlier, most QoS researches in the application layer are related to the

transmission of continuous media streams, and hence are rather media specific and

restrictive in certain application domains [11][12][13][14][15] A multimedia

application typically supports various codecs for media compression such as Motion

JPEG, MPEG-4 and H.264 These codecs present diverse visual-auditory quality to an

end-user by incorporating different compression techniques and compression ratio On

the other hand, different codecs have different emphasis on resource allocation

Theoretically, a highly compressive codec requires more CPU time slice for media

compression and less network bandwidth for data transmission compared with a low

compression ratio codec Hence multimedia applications can choose codecs of

different resource requirements so as to fit into runtime environments of diverse

conditions and resource availability

A multimedia application in networking environments generally will present

delay-sensitive and loss-tolerant characteristics [54] Firstly, most multimedia applications

can cope with certain amount of packet loss depending on the sequence characteristics

and error concealment strategies (e.g packet loss up to 5% or more can be tolerated at

times [55]) Secondly, multimedia applications have stringent delay constraints For

interactive applications such as videoconferencing, delay upper bound is commonly

known as less than 200 milliseconds Comparatively, multimedia streaming

applications can tolerate delay up to 1 or 5 seconds [56] Typically, data packets that

arrive after their display time are discarded at the receiver side or, at best, can be used

for concealing subsequently received multimedia packets The delay-sensitive,

resource-intense and loss-tolerant features of multimedia applications suggest that QoS

management and adaptation can be effective in adjustment of a multimedia

application’s presentation quality in view of runtime dynamics

To assist the design and deployment of multimedia applications, media framework has

been proposed to provide a semantically rich programming environment and facilitate

the access of I/O device and synchronization of different media streams Windows

Media Technology (WMT) [57] and Java Media Framework (JMF) [58] are two

popular media frameworks Platform independent and open source are the advantages

of JMF over WMT JMF enables audio, video and other time-based media to be added

to Java-based applications and can capture, play, stream and transcode multiple media

formats It also supports RTP/RTSP [59][60] in order to interoperate with

standard-based, third-party video streaming servers from, for example, Apple, Sun and Kasenna

Hence, our prototype implementation of QCMF has chosen JMF as the development

platform

2.2.5 Cross-layer QoS Architectures

Layering is a common approach for dealing with the high complexity of QoS

provisioning, so that research issues of each layer can be considered in isolation

Existing QoS literatures mainly deal with QoS provisioning within the context of one

of the individual architecture layers as aforementioned A QoS researcher in this way

would typically focus on one aspect of the QoS provisioning mechanisms for a layer,

neglecting possible related QoS mechanisms in others For example, current end-host

QoS solutions tend to adapt their middleware or applications to the changing network

QoS conditions Thus an ongoing session may have to be aborted when the resource

scarcity in network (e.g., bandwidth shrink) degrades the initially agreed QoS to a

level beyond any end-host adaptation can cope with However, we argue that network

in this case may simply be a better place to exercise QoS adaptation (if the network is

QoS-enabled such as a DiffServ network) so as to prevent the abortion of the session

This example clearly illustrates a serious shortcoming of dealing QoS problems in

isolation, which leads to a less effective end-to-end QoS solution Hence, we assert

that any decent end-to-end QoS solution must consider the interactions of QoS

mechanisms between layers

A number of cross-layer QoS architectures have been proposed to address the QoS

issue by assuming a centralized solution with a single management point and direction

of decision-making[61][62][63] A cross-layer framework jointly analyzes and

optimizes the different strategies available at various system layers (e.g., physical layer,

medium access control (MAC) layer, network/transport layers or applications) For

instance, authors of [64] employ a central coordinator to decide QoS configurations in

three layers of an end-host (i.e., application task, OS scheduler and CPU speed) It

should be noticed that the management scope of most cross-layer proposals are within

one end-host where fine-grained control of different layers can be achieved Although

such a federal solution works for local decision-making within one end-host, it may

not be applicable to end-to-end QoS provisioning in that a local coordinator in one

QoS subsystem is unlikely able to decide QoS configurations and adaptations for

others (such as the network or a remote host) In view of this, QCMF tries to

coordinate QoS efforts from various sub-systems for the benefit of end-to-end

provisioning rather than determining their respective configuration and actions

2.2.6 End-to-end QoS Schemes

As isolated QoS provision may lead to localized QoS solutions which are undesirable

for end-to-end QoS delivery, an overall QoS framework that encompasses QoS

mechanisms of communication components and facilitates implementation that would

harmonize their interactions is ideal for end-to-end QoS transmission Among the few

reported work in the area of integrated end-to-end QoS schemes [65][66][67][68][69],

focuses have been put on connecting respective QoS-flows of each architecture layer

(e.g., interface design, service negotiation protocols [29], specification and translation)

and supporting the underlying enabling mechanisms in each layer For example, the

Enthrone project [30] proposes an integrated management solution which covers an

entire audio-visual service distribution chain, including content generation and

protection, distribution across networks and reception at user terminals Similarly, [70]

proposes a general QoS management framework to select and configure most

appropriate system components according to user requirements and runtime available

resources In [71], authors propose a content-aware bandwidth broker (CABB) to

manage QoS for multimedia applications in a DiffServ environment CABB allocates

network resources to multimedia flows based on client requirements, the adaptability

of the application, and its tolerance to network level parameters such as bandwidth,

delay, and latency Kim et al describes an end-to-end performance simulation model

and methodology for the CDMA 2000 network in [56] The simulator models all

protocol layers from physical to the application layers Details of the packet handling

characteristics of each network element along the end-to-end path are also considered

to compare and measure performance of applications under different settings However,

all these work has overlooked the complexity of end-to-end QoS with respect to

decision-making, especially in the case of QoS adaptation

End-to-end QoS in our view is distributed and heterogeneous in nature; each of its QoS

components may have its own QoS mechanisms and adaptation strategies In this

context, for example, QoS middleware may have its own means of adaptation in case

of QoS violations Meanwhile, adaptive applications may also be able to transform

themselves to cope with runtime changes Things will become more complex if the

network: (1) is also QoS-enabled where diverse service options are of choices, (2)

offers heterogeneous QoS in different network domains, some of which, for example

may employ QoS routing while others may make use of load control or selective

packet dropping techniques [39] Given multiple QoS objectives and QoS service

options on the end-to-end path, a good (coordinated) QoS decision-making will

certainly become more difficult due to an expanded solution space and possible

interactions between QoS options Such a complexity is often not considered in the

aforementioned end-to-end schemes With such a consideration, QCMF is designed to

be an adaptive end-to-end framework with emphasis on system-wide coordinated

adaptation, leveraging on the capabilities of each end-to-end sub-systems

2.3 DY N A M I C P R O T O C O L C O M P O S I T IO N

Dynamic protocol framework (DPF) [46] is a middleware component in QCMF, which

can provide dynamic protocol stack composition at call-setup time and re-composition

(i.e., protocol inserting or swapping) at runtime DPF provides the flexibility of

building a protocol graph of dynamically loaded components supporting media flows,

in a manner similar to other component-based frameworks In the context of

end-to-end provisioning, DPF offers one possible dimension of QoS adaptation within the

communication protocol stack which can supplement current prevailing QoS solutions

at application or network level

In DPF, protocol components need not to be bound at design time, which provides the

flexibility in composition of protocol stacks Instead of specifying the name of a

protocol, applications now can specify their desired QoS properties For example, a

multimedia application may request to reserve resources before its session starts At

build-time, available protocol services that match this requirement (e.g., RSVP or

other signaling protocol with similar resource reservation capability) will be selected

Such an approach eases the deployment of multimedia applications in heterogeneous

network environments in that if a target protocol is not available, other protocols of

similar functions can be selected so that the end-to-end delivery will not fail (e.g.,

[72]) The flexible composition of protocol stacks also facilitates the QoS adaptation

process For instance, two video codecs may present similar presentation quality to an

end-user, but at different compression rate and hence each demands for different

amount of network bandwidth In the case of QCMF/DPF where codec names need not

be specified by multimedia applications (instead, media quality such as medium or

high should be specified), a codec that requires more bandwidth may be replaced at

runtime with another one that consumes less bandwidth in case of network congestion

To ensure a consistent description of all end-to-end QoS entities, we have designed a

semantic scheme for modeling and processing of protocol stacks, which is presented in

Section 4.3 The semantic model of communication protocols and protocol instances

are also illustrated in Appendix B The integrity of protocol and protocol stack

configuration are ensured with sets of dependencies and supported media formats

primitives to be defined by protocol developers Service dependencies and media

format compatibility are checked at stages of both build-time configuration and

runtime reconfiguration to ensure the correctness of a protocol stack

The complex QoS problem has led researchers to focus on different aspects of QoS

provisioning in a fashion similar to the layered approach in network systems design

This has resulted in many rigid QoS solutions each addressing one or very few aspects

of the problems with respect to a set of application scenarios, middleware, or networks

These silos of solutions are either too difficult to integrate, or if doable, often would

lead to overall inefficiency due to poor coordination between respective QoS

sub-systems Hence, we believe that any satisfactory end-to-end QoS solution must

consider the coordination of QoS mechanisms between QoS sub-systems (such as

those in end-hosts and networks) and manage them in a cohesive and coordinated

fashion Through comparison and discussion, we have found that most existing

end-to-end QoS schemes focus primarily on the configuration issues such as interface design

and QoS-flow management, which is essential, but not sufficient for meeting

performance requirements of multimedia applications Motivated by these

observations, we propose our ideas of end-to-end QoS collaboration by

accommodating and coordinating exiting QoS architectures in applications,

middleware and networks Details of our approach will be discussed in the following

chapters

Chapter 3 The QoS Coordination and Management Framework

T H E Q O S C O O R D I N A T I O N A N D

M A N A G E M E N T F R A M E W O R K

3

This chapter gives a high level overview of the architecture and management functions

of our QoS coordination and management framework (QCMF) We first present a

reference model for end-to-end QoS provisioning and discuss our design philosophy

and relevant QoS concepts Subsequently, we introduce the system architecture and

management functionalities of QCMF, whereby detailed description of our research

will be presented in the next few chapters

Figure 3-1: Reference model for end-to-end QoS provisioning and coordination

To deal with the complexity of end-to-end QoS provisioning, we introduce a reference

model to guide the design of our end-to-end QCMF framework as shown in Figure 3-1

This model outlines relevant concepts and procedures for end-to-end QoS provisioning

which can be analyzed from both architecture and management dimensions

From the architecture perspective, the model will yield the identification of several

abstracted QoS layers and their corresponding roles in end-to-end QoS delivery:

• System QoS includes efforts from a host’s OS and the network which provides

basic data transmission support between end-hosts Native packet level QoS

support can be offered if the OS and the underlying network are QoS-enabled

In addition to data link or MAC level QoS provisioning [73], research concerns

in this area have also included network communication level load balancing and

fairness issues [74][75]

• Middleware QoS offers a rich set of services for the configuration and

management of the transmission quality (e.g., buffer management, flow

synchronization and QoS-based handover) outside the kernel space of an

end-host [21][22][23][24] Middleware QoS solutions are likely to be independent of

the network and OS platforms and hence are able to work over heterogeneous

network environments

• Application QoS refers to the ability of multimedia applications to

self-configure and respond to the changes of runtime operating conditions or user

requirements As discussed, such abilities are commonly related to the

transmission and performance tuning of particular continuous media streams

such as audio and video (e.g variable bit rate codec or layered encoded audio

and video) [14][15] Hence QoS solutions at application level are rather media

specific and restrictive to a certain application domain

For over a decade, researchers have proposed various QoS solutions which according

to their places of research interest, can be summarized into one of the above categories

As has been explained in Chapter 2, QoS researchers in this way would typically focus

on their own domains of QoS provisioning while neglect (possible related) QoS

mechanisms in others Such a layered QoS research leads to an independent and local

optimized implementation, but would often result in sub-optimal end-to-end

performance In this sense, an overall QoS framework that encompasses QoS

mechanisms of various layers is essential for the satisfaction of an end-user

Furthermore, end-to-end QoS in our view would be distributed and heterogeneous in

that each QoS layer (subsystem) may have its own provisioning mechanisms and

adaptation strategies An end-to-end QoS framework thus should take into

consideration the characteristics and restrictions of each end-to-end QoS sub-system

(e.g., QoS layer) so that a sound overall adaptation solution can be identified among

multiple available end-to-end choices at runtime

Based on the above design philosophy, we have arrived at the design of QCMF as an

adaptive end-to-end QoS coordination and management framework Our solution

embraces existing and new QoS mechanisms at three entity levels: the network level,

the middleware level and the application level We treat each of these QoS

sub-systems as a QoS component in our end-to-end framework and try to devise an

effective platform and methods for accommodating and supporting interactions and

dynamic adaptations among them In this context, we are not participating in the

performance tuning or enhancement in QoS mechanisms of a particular QoS

component Instead, our focus is to provide a management platform for harmonizing